# Connecting your AI assistant

> How to connect Claude, ChatGPT or another assistant to your mailboxes, and what it can reach.

- Tool: Warsha Emails
- Section: Start here
- Applies to version 2026-10-16
- Last updated: 2026-10-16
- Source: https://warshalabs.com

---

## What are you trying to do?

Connect your AI assistant to your mailboxes so it can read your mail, write drafts and request sending — without sending anything by itself, and without ever seeing a mailbox password.

## Steps

1. **Add at least one mailbox** and make sure its test passed. If you connect the assistant first, it still sees every tool, and they start working the moment a mailbox is added.

2. **Open “Connect the assistant”** and copy **“The connection address for every app”**.

3. **In Claude (web and app):** Settings → Connectors → Add custom connector → paste the address → Connect.

4. **In ChatGPT:** Settings → Connectors (turn on developer mode if asked) → Create → paste the address → choose direct sign-in for authentication.

5. **Allow the connection.** A Warsha Emails page opens: sign in, choose the workspace, the mailboxes and the duration, then “Allow”.

6. **Start.** Ask the assistant, for example, to summarise today's mail. It reads the working rules first on every new connection.

**A second way, for apps that ask for a key** (such as Claude Code and Cursor): in “Connect the assistant” choose “New key”, name it, set its duration, the mailboxes or folders it can reach, and “Read only” if you wish. Then copy the key at once — **it is never shown again**.

## What can go wrong

**The assistant says a tool is not allowed.** Its permissions are your permissions in this connection only: the key may be read-only, limited to some mailboxes, or your plan may not allow the action. The assistant explains the reason and how to get it.

**You changed the allowed mailboxes and the assistant does not see it.** Choose “Refresh” beside it in “Active connections”, and ask the assistant to read the working rules again.

**A “more than 40 tools” warning in Cursor.** Some versions warn at that number. Use a read-only key, or turn off some tools in the app.

## Something you won't expect

**The assistant never sends a single message by itself.** Every send becomes a request waiting for your approval, and “send it” in the chat is not approval. Nor can it change a password, delete your account, pay on your behalf or approve for you — even if you ask it to.

The assistant's provider (Anthropic, OpenAI or another) processes what the assistant reads from your mail under its own policy. Disconnecting stops everything after it at once, but does not recall what was processed before.